name: Replicate Vocabulary
description: >-
  Vocabulary and taxonomy for the Replicate machine learning platform API.
  Covers model inference, deployment, training, and platform concepts.
version: "1.0.0"
created: "2026-05-02"
modified: "2026-05-02"

categories:
  - name: Core Concepts
    description: Fundamental concepts of the Replicate platform.
    terms:
      - term: Prediction
        definition: >-
          A single inference run against a machine learning model. A prediction
          takes input, runs the model, and produces output. Predictions can be
          synchronous (wait mode) or asynchronous (webhook-based).
        related: [Model, Version, Deployment]

      - term: Model
        definition: >-
          A machine learning model hosted on Replicate. Models are identified
          by owner/name (e.g., stability-ai/sdxl) and can be public or private.
          Each model has one or more versions.
        related: [Version, Prediction, Training]

      - term: Version
        definition: >-
          An immutable snapshot of a model at a specific point in time.
          Identified by a SHA256 hash. Predictions run against a specific
          version for reproducibility.
        related: [Model, Prediction, Cog]

      - term: Deployment
        definition: >-
          A production-ready, auto-scaling instance of a model version.
          Deployments provide lower latency and guaranteed capacity.
          Configured with hardware type and min/max instance counts.
        related: [Model, Version, Hardware]

      - term: Training
        definition: >-
          A fine-tuning job that customizes a base model with new data.
          Trainings create new model versions. Replicate supports fine-tuning
          with LoRA and other parameter-efficient methods.
        related: [Model, Version, Prediction]

  - name: Lifecycle States
    description: Status values for predictions and trainings.
    terms:
      - term: starting
        definition: "The prediction or training has been created but not yet started."
        applies_to: [Prediction, Training]

      - term: processing
        definition: "The prediction or training is currently running."
        applies_to: [Prediction, Training]

      - term: succeeded
        definition: "The prediction or training completed successfully with output."
        applies_to: [Prediction, Training]

      - term: failed
        definition: "The prediction or training failed with an error."
        applies_to: [Prediction, Training]

      - term: canceled
        definition: "The prediction or training was canceled by the user."
        applies_to: [Prediction, Training]

  - name: Infrastructure
    description: Hardware and deployment infrastructure concepts.
    terms:
      - term: Hardware
        definition: >-
          The GPU hardware type used to run a model. Options include T4, A40,
          A100, H100 GPUs at various price points. Hardware is specified when
          creating models and deployments.
        related: [Model, Deployment]

      - term: Cog
        definition: >-
          Replicate's open-source tool for packaging machine learning models
          into containers. Models are built with Cog and pushed to Replicate.
        url: https://github.com/replicate/cog
        related: [Model, Version]

      - term: Auto-scaling
        definition: >-
          The ability for deployments to automatically add or remove instances
          based on demand. Configured via min_instances and max_instances.
        related: [Deployment]

  - name: Integration Patterns
    description: Common patterns for integrating with the Replicate API.
    terms:
      - term: Webhook
        definition: >-
          An HTTPS callback URL that Replicate calls when a prediction or
          training has new output. Enables asynchronous notification without
          polling.
        related: [Prediction, Training]

      - term: Webhook Events Filter
        definition: >-
          A filter that controls which prediction lifecycle events trigger
          webhook calls. Events include: start, output, logs, completed.
        related: [Webhook, Prediction]

      - term: Wait Mode
        definition: >-
          A synchronous prediction mode where the API request stays open until
          the model finishes. Enabled via the Prefer header (e.g., wait=5).
          Useful for fast models.
        related: [Prediction]

      - term: Streaming
        definition: >-
          Real-time output streaming via Server-Sent Events (SSE). The prediction
          URL contains a stream entry for models that support streaming output.
        related: [Prediction]

      - term: Signing Secret
        definition: >-
          A shared secret used to verify that incoming webhook payloads
          are genuinely from Replicate. Retrieved via /webhooks/default/secret.
        related: [Webhook]

  - name: Collections
    description: Model discovery and curation concepts.
    terms:
      - term: Collection
        definition: >-
          A curated group of models organized by category or use case.
          Examples include super-resolution, text-to-speech, image-restoration.
        related: [Model]

      - term: Slug
        definition: >-
          A URL-friendly identifier for a collection (e.g., super-resolution,
          text-to-image). Used in API paths to identify collections.
        related: [Collection]